ChatGPT 4: The lawyer who knows how to lie perfectly

ChatGPT 4: The lawyer who knows how to lie perfectly

Share post

The update for ChatGPT to version 4 was eagerly awaited and, according to the manufacturer Open AI, is said to represent the “most advanced” AI technology and even passed a bar exam with top marks in a test. However, the only thing that ChatGPT can recreate are lies that are hardly recognizable in the factual texts, according to Sophos security expert Chester Wisniewski in an interview.

How does a representative of the IT security industry see the rapid development in chat AI and how does he classify the resulting opportunities and possible disadvantages? Chester Wisniewski, Principal Researcher at Sophos, answered some of the questions. His conclusion: the benefit probably outweighs the harm.

What are the differences between GPT 3 and the new GPT 4?

A whole lot has changed and improved, but not everything is known. OpenAI is keeping its cards a little tighter with this release than in the past, so not all improvements have been publicized. On the whole, it seems smarter, more accurate, and more capable than previous versions, which should make the reactions even more realistic and competent. It's still important to remember that while it's less likely to spread false information, it's still a pretty good liar.

Is it ethical to make such a generative, pre-trained transformer publicly available?

He is just as available to the bad guys as to the good guys. Technological progress is like a genie in a bottle. You can't just put it away when it gets uncomfortable. Having tools like ChatGPT available with the protections OpenAI is trying to enforce is unlikely to do more harm than good. While criminal abuse of AI is inevitable, the good that can come from responsible use of these tools will likely far outweigh any abuse.

It seems that there are already numerous examples of prompt engineering breaking the system's guard rails, and there is at least one example of indirect prompt injection. Do you think it will ever be possible to create a GPT model that cannot be abused?

🔎 Interview: Sophos security expert Chester Wisniewski (Image: Sophos).

No. A more detailed answer would probably be more welcome now, but any system designed to prevent abuse but also designed to allow the system to work autonomously will most likely always be circumventable. But as OpenAI and other AI researchers learn more about how people circumvent their defenses, this becomes more difficult. I hope and expect that it will take significantly more dexterity on the part of cybercrime in the future. However, it will still be possible to compromise the system.

Is ChatGPT creating anything new, or just doing what it's been told? What if the model has learned or been taught inaccuracies?

The only thing ChatGPT can recreate is lies. It's a pretty convincing liar that rivals the best narcissists I know. The model has certainly also been taught things that are inaccurate or untrue, and it is almost impossible to predict when it will incorporate these factual inaccuracies into its answers.

Big AI will be owned by Big Tech. While the personal and sensitive company data contained in the prompts may not feed into the particular AI model, it will certainly be available to the big tech companies. And the business model of big tech owners often relies on selling such data to the highest bidder. Can something be done about the privacy aspects of GPT-style public AI?

For most users of AI systems today, there is little choice but to trust the promises and license agreements they agree to in order to use models like ChatGPT. However, many of these models are developing incredibly efficiently, so there is no reason why they cannot run on our PCs, laptops or even phones in the future. Training the models is computationally intensive, but executing it is not. Older models may be offered as open source or even crowdsourced so that they are available to anyone who wants to use them. The most cutting-edge models will likely remain in the hands of the big tech giants, but models that are more than "good enough" could run on an iPhone in your pocket if you wish.

More at Sophos.com

 


About Sophos

More than 100 million users in 150 countries trust Sophos. We offer the best protection against complex IT threats and data loss. Our comprehensive security solutions are easy to deploy, use and manage. They offer the lowest total cost of ownership in the industry. Sophos offers award-winning encryption solutions, security solutions for endpoints, networks, mobile devices, email and the web. In addition, there is support from SophosLabs, our worldwide network of our own analysis centers. The Sophos headquarters are in Boston, USA and Oxford, UK.


 

Matching articles on the topic

Wireless security for OT and IoT environments

Wireless devices are becoming more and more common. This increases the number of access points through which attackers can penetrate networks. A new ➡ Read more

Professional cybersecurity for SMEs

Managed detection and response (MDR) for SMEs 24/7, 365 days a year. The IT security manufacturer ESET has expanded its offering ➡ Read more

Qakbot remains dangerous

Sophos X-Ops has discovered and analyzed a new variant of the Qakbot malware. These cases first appeared in mid-December and they ➡ Read more

Prevent malicious software from starting

A cyber protection provider has added a new feature to its security platform. It improves cybersecurity by preventing the launch of malicious or ➡ Read more

Pikabot: camouflage and deceive

Pikabot is a sophisticated and modular backdoor Trojan that first appeared in early 2023. His most notable quality lies in ability ➡ Read more

Ransomware-resistant WORM archives for data backup 

A data archive is a must for every company. Few people know: An active WORM archive can help to streamline data backup, ➡ Read more

LockBit is alive

A few days ago, international law enforcement authorities scored a decisive blow against Lockbit. According to a comment from Chester Wisniewski, Director, Global ➡ Read more

Danger of election manipulation through cyber attacks

Cyberattackers are attempting to influence elections around the world using generative AI technology. The latest findings from the Global Threat Report ➡ Read more